This invention relates to systems and methods for acquiring biometric and other imagery, biometric acquisition, identification, fraud detection, and security systems and methods, particularly biometric systems and methods which employ iris recognition. More particularly the invention relates to systems and methods for acquiring iris data for iris recognition.
Iris recognition systems have been in use for some time. The acquisition of images suitable for iris recognition is inherently a challenging problem. The performance of recognition algorithms depends on the quality, i.e., sharpness and contrast, of the image of the iris of the subject who is to be identified. This is due to many reasons. As an example, the iris itself is relatively small (approximately 1 cm in diameter) and it is often required to observe it from a great distance in order to avoid constraining the position of the subject or when the subject is walking or riding. This results in a small field of view and also a small depth of field. As a second example, it is generally difficult for the adult or child subject to stay absolutely still. As a third example, the subject may blink involuntarily or drop or swivel their head momentarily to check on the whereabouts of luggage.
In biometric identification applications, due to unconstrained motion of cooperative or non-compliant subject, it has been very difficult to acquire iris images with sufficient quality for recognition and identification processing. For example, iris acquisition systems typically check whether the quality of an acquired image exceeds a threshold. Many methods of assessing quality have been developed, such as those based on a measurement of focus such as those disclosed in U.S. Pat. No. 6,753,919,
The problem with this approach is that if the acquired image quality does not exceed the threshold, then the data is not acquired, despite the fact that there may never be another opportunity to acquire data from that subject again. More specifically, in the case of unconstrained users or non-cooperative subjects, it may be impossible to have the subject position themselves or wait until the acquired image data exceeds the quality threshold. For example, the subject may be distracted with their head turning in various directions, or they may be in the process of performing another task, such as boarding a bus, so that the opportunity to acquire data from them has already come and gone. More specifically, prior iris data acquisition systems have typically been designed to explicitly avoid capturing lower quality data with an emphasis on waiting or constraining the user such that only highest quality data is acquired. We have determined that even a lower quality iris image (blurred, for example) can still contain substantial evidence for matching, albeit not with the precision of a high quality iris image. However, we still wish to acquire high quality data when it is possible to do so. In another example of prior systems, for example those disclosed in U.S. Pat. No. 5,151,583, autofocus routines are used to attempt to obtain high quality iris images. However, autofocus routines cause lag times and inaccuracy, resulting in poor quality or even non-existent imaging. Other systems, such as the ones disclosed in U.S. Pat. No. 6,753,919 by Daugman, use sensors to assist a subject in aligning and focusing a handheld video camera.
Most if not all automatic focus systems work by acquiring an image of the scene, processing the image to recover a measure of focus, using that measure of focus to move a lens-focus actuator, and then repeating the steps of image acquisition, processing and actuation many times until it is determined in the processing step that focus has been reached. In most iris recognition systems autofocus never is able to catch up with the actual position of the subject unless the subject is relatively stationary, due to the unusually low depth of field in iris recognition, as well as the requirement that the focus has to be on the iris (as opposed to the nose for example).
Because of the time delays involved in acquiring an image, processing the image, and mechanical actuation, it is impossible for auto-focus algorithms to respond instantaneously. Moreover, as the depth of field reduces, as is typically the case in iris recognition, where the object is small and is typically observed at high magnification, it becomes more difficult for auto-focus algorithms to be successful because any error in the auto-focus position is much more apparent in the imagery since the depth of field is small.
It is much more difficult for auto-focus to acquire in-focus imagery of a subject who is moving even slightly (fractions of an inch).
In the case of a person moving even slightly because there is a finite control loop time for standard auto-focus to actuate, it can be shown that if a component of the person's motion is high frequency and above the control loop response time, then the auto-focus will never be able to converge and acquire an in-focus image of the person. The auto-focus will be continually “hunting” for a focused image and will always lag the motion of the subject. The result is that the subject has to be rock solid and still when standard auto-focus is used, and this was the state of the art in iris recognition before the present invention.
Prior attempts to solve these autofocus problems use the same closed loop approach but assume a subject is moving in a straight line and then use the image measurements to try and predict where the person will be in the next frame. This approach is not very robust and also fails for random movement that subjects often have. Other auto-focus systems use different ways of computing focus measures in the scene in one or more regions to compute the most accurate focus score. When a subject is moving with frequencies that are beyond the control loop of an auto-focus algorithm auto-focus algorithms are unable to catch up to the person's motion and acquire a good image of the person.
Martin, et al., US Pat. Pub. 2008/0075335, disclose a biometric image selection method which reduces the rate of non-exploitable images which are supplied to an analysis and identification processing module using sharpness and contrast criteria. In some embodiments Martin et al. locate a pattern in each image of a sequence of images, estimate the speed of displacement of the pattern between two successive images in the sequence, and select images for which the estimated speed of displacement of the pattern is lower than a speed threshold. Martin et al. disclosed embodiments wherein two selection modules are provided, the first being a quick selection module and the second being a pupil tracking module, rejecting an image if it is below a contrast or sharpness threshold. The selection module in some embodiments selects images having the highest sharpness and/or contrast out of the images stored. Martin et al do not disclose a system or method for acquiring the series of images, nor do they disclose storing only images having higher quality than previously stored images and removing the lesser quality image from memory storage.